# Brutha by Ergonsun

> Brutha is a local-first, self-hosted persistent memory system, portable Character engine, and optional Persistent Code Knowledge Graph for AI assistants and multi-agent workflows. Built by Ergonsun (`https://brutha.ergonsun.ai`), Brutha runs with zero LLM calls in the memory path, $0/query usage cost, and zero cloud egress.

## Product Overview

Brutha solves context loss, factual drift, and cross-session amnesia across AI assistants. Unlike cloud-hosted AI memory wrappers that send user conversations to external LLM APIs on every save and search, Brutha executes all memory ingestion, hybrid retrieval, semantic contradiction resolution, synthesis, and codebase mapping locally on the user's machine (Brutha Solo) or on the organization's own Postgres infrastructure (Brutha Team & Enterprise).

Brutha connects via the Model Context Protocol (MCP) to Claude Code, Claude Desktop, OpenAI Codex CLI, Google Gemini CLI, Cursor, Windsurf, Zed, Kilo Code, OpenCode, Local Llama / Ollama setups, and custom agent frameworks.

## Core Capabilities

1. **Zero-LLM-Call Persistent Memory ($0/query)**:
   - Hybrid semantic + lexical retrieval with sub-50ms p95 latency (`26.9 ms` p95 on `memory_recall`, `23.0 ms` p95 on `memory_search`, `37.7 ms` p95 on `memory_save`).
   - Snippet-first token budgeting: returns compact snippets first and expands full memory records only on demand (`memory_expand`), saving over 90% of context tokens.
   - Provenance and epistemic trust: tracks `user_stated`, `tool_verified`, `external_doc`, and `agent_inferred` sources, plus automatic aging (`hot` -> `warm` -> `cold`) without destructive deletion.

2. **Git-Style Semantic Conflict & Contradiction Resolution**:
   - Detects duplicate discoveries and increments corroboration counts instead of polluting storage.
   - Automatically detects semantic contradictions in real prose (including polarity inversions, numeric/unit shifts, status changes, and multi-sentence factual reversals) and links superseding facts (`supersedes`, `caused_by`, `relates_to`).

3. **Portable AI Characters (`character_activate`, `character_save`)**:
   - Stores durable role personas, operating rules, preferred model hints, and domain workflows in the memory database so any MCP client can adopt the exact same specialist persona at session start.

4. **Persistent Code Knowledge Graph (19 Languages)**:
   - Indexes repositories into a persistent symbol and dependency graph (`code_index_repo`, `code_map`, `code_search`, `code_symbol`, `code_neighbors`, `code_trace`, `code_impact`, `code_architecture`).
   - Anchors architectural decisions and bug memories directly to code symbols (`code_memory_link`, `code_memory_context`) and detects symbol drift across commits.

## Reproducible Benchmarks

Public machine-readable evaluation artifacts are hosted directly on `brutha.ergonsun.ai`:

- **LongMemEval Tier-1 (48-Question Retrieval & Reader Curve)**:
  - Report: `https://brutha.ergonsun.ai/benchmarks/longmemeval-tier1.json`
  - **92.8% Recall@5** and **94.8% End-to-End Judge Accuracy** at top-k=5 (`0.849` MRR, `0.869` NDCG@5), using ~1,978 prompt tokens vs. ~115,000 tokens for full-history stuffing (**58x fewer prompt tokens**, **0 LLM calls on ingest**).
- **106-Pair Real-Prose Contradiction Suite**:
  - Report: `https://brutha.ergonsun.ai/benchmarks/prose-contradiction-106.json`
  - **89.3% F1** (**92.0% Precision**, **86.8% Recall**) across 106 adversarial real-prose pairs at sub-25ms local inference.
- **15-Task Code-Graph & Drift ROI Benchmark**:
  - Report: `https://brutha.ergonsun.ai/benchmarks/code-graph-roi.json`
  - **29.4x Token ROI** (~1,420 tokens vs. ~41,800 tokens for raw grep/file reads), **3.4x fewer tool calls** (`2.1` vs `7.2`), and **100% (15/15) symbol-drift detection**.

## Editions & Pricing

- **Brutha Solo (`/` and `/download`)**:
  - **Price**: **$0 (Free forever)** for individuals across all professions, with optional Stripe supporter contributions.
  - **Storage**: Single local binary (`brutha-mcp`) backed by local SQLite (default) or single-owner local Postgres.
  - **Tools**: 40 MCP tools (29 core memory, synthesis, lifecycle, and Character tools + 11 Code Knowledge Graph tools).
- **Brutha Team Edition (`/enterprise`)**:
  - **Price**: **$29/seat/month** billed annually, or **$35/seat/month** billed monthly (5 to 25 seats).
  - **Architecture**: Self-hosted `brutha-enterprise-mcp` on your team's Postgres with 1 shared team schema, private vs. shared visibility, and a shared Team Character cast.
- **Brutha Enterprise Edition (`/enterprise`)**:
  - **Price**: **$59/seat/month** billed annually, or custom annual agreement (25+ seats, unlimited teams).
  - **Architecture**: Physical per-team Postgres schema isolation (`team_<slug>`), four-tier visibility (`private`, `team`, `cross_team`, `enterprise`), PAT authentication, and full auditability.

## Key Links

- Homepage (Brutha Solo, Architecture & Benchmarks): https://brutha.ergonsun.ai/
- Teams & Enterprise (Architecture, Security & Pricing): https://brutha.ergonsun.ai/enterprise
- Documentation Hub (Solo & Enterprise Task Guides): https://brutha.ergonsun.ai/docs
- Download Brutha Solo: https://brutha.ergonsun.ai/download
- Release Notes (v1.0.0 Solo & Enterprise): https://brutha.ergonsun.ai/releases
- Extended AI Technical Reference: https://brutha.ergonsun.ai/llms-full.txt